Abstract
Natural Language Understanding (NLU) has long been a cornerstone task within the field ofNatural Language Processing (NLP), particularly due to the implicitness, ambiguity, and com-
positional complexity of natural language. Language models often struggle with reasoning over
latent semantics and underspecified discourse structures. In this dissertation, we introduce Dense
Paraphrasing (DP), a linguistically motivated textual enrichment framework designed to surface
implicit meaning and enhance compositional reasoning. DP transforms raw textual inputs into se-
mantically enriched paraphrases by recovering elided arguments, clarifying event dynamics, and
aligning referential structures. In this dissertation, we demonstrate the effectiveness of DP across
four representative tasks: (1) improving semantic grounding in sentence similarity, (2) enabling
multimodal grounding in situated dialogue, (3) eliciting implicit semantics via question answer-
ing, and (4) tracking event-based transformations through coreference and subevent modeling. The
experiments conducted in this research consistently show improvements in model performance, in-
terpretability, and contextual understanding.